Fast Visual Feature Selection and Tracking in a Hybrid Reconfigurable Architecture

1
Fast Visual Feature Selection and Tracking in a
Hybrid Reconfigurable Architecture
Center for Embedded Networked Sensing
Alessandro Bissacco Jason Meltzer
Soheil Ghiasi Stefano Soatto UCLA Vision
Lab - http//vision.ucla.edu ER Systems
Lab - http//er.cs.ucla.edu
Visual Feature Selection and Tracking in Video
Sequences
Selection Find salient points in the first image
frame
Tracking Estimate the positions of selected
points in the next frames
Problem Description Real-time Selection and
Tracking on 640x480 videos at 60Hz

• Real-time Selection and Tracking are
  computationally expensive.
• With state-of-the-art software implementations,
  on a 3.2 GHz Pentium IV machine
• Selection more than 100 ms on 640x480 images
• Tracking 60Hz possible only for about 100
  features

Proposed Solution Hybrid Reconfigurable
Architecture for Fast Selection and Tracking
 

• Feature selection is implemented in FPGA
  hardware. The algorithm is an adaptation of the
  Lucas-Kanade feature selector 1 to integer
  arithmetic.
• Kalman filtering is used to predict feature
  locations in the next frame.
• A Nearest Neighbor data association technique 2
  is used to find correspondences.
• Normalized Cross Correlation is applied to prune
  incorrect correspondences.

• Hardware
• Xilinx FPGA WILDSTAR/PCI board at 100 MHz with 1
  million gates, 1.6 Gb/s I/O bandwidth, 6.4 Gb/s
  memory bandwidth.
• Pentium IV 3.2GHz, 1 Gb RAM.
• Results
• Selection about 0.16 ms on 640x480 images.
• Tracking table below.

UCLA UCR Caltech USC CSU JPL UC
Merced